Simulation-Based Parameter Optimization for Self-adaptive HPL on Parallel Systems

Computational benchmarks are essential for dependable systems, applications, and technologies across multiple domains. However, traditional benchmarks such as High-Performance Linpack (HPL) require parameters to be set to reflect the characteristics of the system under test, and the adoption of empi...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:International journal of parallel programming Ročník 53; číslo 4; s. 24
Hlavní autoři: Rista, Cassiano, Teixeira, Marcelo, Fonseca, Mauro
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York Springer US 01.08.2025
Springer Nature B.V
Témata:
Benchmarks
Computer Science
Parameters
Processor Architectures
Simulation
Simulation models
Software Engineering/Programming and Operating Systems
Stress analysis
Theory of Computation
Simulation
Empirical development
Benchmarks
Autonomous computing
Modeling
Model-driven development
Reconfiguration
ISSN:0885-7458, 1573-7640
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Computational benchmarks are essential for dependable systems, applications, and technologies across multiple domains. However, traditional benchmarks such as High-Performance Linpack (HPL) require parameters to be set to reflect the characteristics of the system under test, and the adoption of empirical fine-tuning strategies is inevitable. As an alternative, adopting simulation models allows the representation of the stochastic nature of benchmark events and provides in-advance estimations instead of stress analysis. This paper extends the HPL benchmark with self-adaptation capabilities (SA-HPL) and proposes a simulation model that enables the automatic configuration of the SA-HPL parameters, allowing the benchmark to adapt effectively, quickly, and autonomously to various scheduling strategies, including static and on-demand schedulers. Experiments conducted over real data show reasonable accuracy between estimated and measured performance. The observed Mean Absolute Percentage Error in throughput was 1.26% for the Server and 3.76% for the Workstation experiments, while for efficiency, they were 1.15% and 3.52%, respectively. These results suggest significant savings in time, effort, and costs.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0885-7458
1573-7640
DOI:10.1007/s10766-025-00803-5